Research Engineer at LINE WORKS. Working on Research in AI(NLP, Vision, Audio, ...), HPC, Physics, and their combination. ex-OMRON SINIC X. Ph.D. from UTokyo.
We sped up materials simulations by 36x with Claude Fable 5.2.
In our early testing, Claude Fable 5.2 sped up DFT and DFPT codepaths in Quantum Espresso by 36x and 9x respectively. It did this while running largely in an autonomous loop, checking its work against our verification sets.
Quantum Espresso is the most widely used open-source codebase for DFT, with roughly 600,000 lines of code written in Fortran.
Thread 🧵
Hi all, I have a new result I'm excited to share. It turns out that *all existing implementations* of softmax and cross-entropy (that's JAX, PyTorch, textbooks, etc) have a subtle bug --- two bugs, actually. A simple rewrite fixes both, and eliminates "slingshots" when grokking: